<![CDATA[The rise of political discourse on Indonesian social media platforms such as X (formerly Twitter) creates opportunities and challenges for policymakers. Existing sentiment analysis methods often fail to handle informal language, slang, and sarcasm, leading to frequent misclassification that may misguide governance decisions. This study aims to establish the first benchmark for three-class sentiment analysis (positive, neutral, negative) in Indonesian political discourse using a Long Short-Term Memory (LSTM) model with culture-specific preprocessing. A dataset of 1,002 tweets on the performance of the Governor of West Java (Feb–May 2025) was collected, normalized for slang and typos, and enriched with a political lexicon. Manual annotation achieved high agreement (κ = 0.82). An LSTM model with 128 units and 30% dropout was trained and evaluated. Results show 95.88% training accuracy but only 36.32% validation accuracy, indicating severe overfitting. Misclassifications (42%) mainly stemmed from sarcasm and contextual ambiguity, with the lowest precision in the positive class (31%). The study contributes by (1) providing the first benchmark for Indonesian political sentiment, (2) demonstrating the value of culture-specific preprocessing, and (3) offering policy insights into latent dissatisfaction hidden in neutral tweets. Limitations include small dataset size and lack of sarcasm-aware mechanisms, suggesting future exploration of hybrid and transformer-based models.]]>
